/finetuning

Evaluation Methodology

Principles

1. Always establish a baseline: Evaluate the base (un-tuned) model first. Without a baseline, you can't measure improvement.
2. Use a held-out test set: Never evaluate on training or validation data. The model has seen those.
3. Use the same test set for every model: This is the only way to compare results fairly.
4. Use task-specific graders: Built-in generic evaluators (Coherence, Fluency) measure general quality and won't detect fine-tuning improvements. Use custom graders (Python, score model, string check) for task-specific evaluation.
5. Measure cost alongside accuracy: Report completion tokens per response when comparing models or checkpoints. A model that achieves the same accuracy with fewer tokens is strictly better — cheaper inference and lower latency.

Two-Layer Evaluation Strategy

Use the Azure AI Evaluation SDK (azure-ai-evaluation) for all evaluation.

Generic built-in evaluators (Coherence, Fluency, TaskAdherence) are guardrails, not metrics — they often show no difference between base and fine-tuned models even when domain-specific evaluation reveals clear improvement.

Custom Graders (Primary FT Evaluation)

1. Score Model Grader (LLM judge with task-specific rubric)

Best for: subjective tasks (summarization, alignment, style).

from azure.ai.evaluation import AzureOpenAIScoreModelGrader

summarization_grader = AzureOpenAIScoreModelGrader(
    model_config=model_config,
    name="summarization_quality",
    prompt="""Rate this news summary on a scale of 1-5.

Article: {{item.article}}
Summary: {{sample.output_text}}

Criteria:
- Captures ALL key facts (who, what, when, where)
- No hallucinated information not in the article
- Concise (under 3 sentences)

Score 1: Missing key facts or hallucinations
Score 3: Captures main point but misses details
Score 5: Perfect summary — all facts, no extras, concise

Return ONLY a number 1-5.""",
    output_type="numeric",
    pass_threshold=3,
)

2. Python Grader (programmatic/exact-match evaluation)

Best for: code generation, math, entity extraction, structured output.

from azure.ai.evaluation import AzureOpenAIPythonGrader

entity_grader = AzureOpenAIPythonGrader(
    name="entity_extraction_accuracy",
    source="""
import json

def grade(item, sample):
    try:
        extracted = json.loads(sample["output_text"])
        reference = json.loads(item["ground_truth"])
    except (json.JSONDecodeError, KeyError):
        return {"score": 0, "reason": "Invalid JSON output"}

    required_keys = ["people", "organizations", "locations", "dates"]
    missing = [k for k in required_keys if k not in extracted]
    if missing:
        return {"score": 0.5, "reason": f"Missing keys: {missing}"}

    total, matched = 0, 0
    for key in required_keys:
        ref_set = set(str(v).lower() for v in reference.get(key, []))
        ext_set = set(str(v).lower() for v in extracted.get(key, []))
        total += len(ref_set)
        matched += len(ref_set & ext_set)

    score = matched / total if total > 0 else 1.0
    return {"score": score, "reason": f"{matched}/{total} entities matched"}
""",
    pass_threshold=0.7,
)

3. String Check Grader (pattern matching)

Best for: classification, format compliance, tool calling format.

from azure.ai.evaluation import AzureOpenAIStringCheckGrader

tool_format_grader = AzureOpenAIStringCheckGrader(
    name="tool_call_format",
    input="{{sample.output_text}}",
    operation="like",          # or "eq", "starts_with", "contains"
    reference="function_call",
    pass_threshold=1,
)

classification_grader = AzureOpenAIStringCheckGrader(
    name="classification_accuracy",
    input="{{sample.output_text}}",
    operation="eq",
    reference="{{item.expected_label}}",
    pass_threshold=1,
)

Running an Evaluation

The evaluate() function runs multiple graders over an entire dataset:

from azure.ai.evaluation import evaluate, F1ScoreEvaluator

result = evaluate(
    data="eval_data.jsonl",
    evaluators={
        "task_grader": my_custom_score_grader,   # primary
        "f1": F1ScoreEvaluator(),                 # token overlap
    },
    output_path="./eval_results.json",
)

for metric, value in result["metrics"].items():
    print(f"{metric}: {value}")

Test Set Design

• Size: 30–100 examples is sufficient.
• Diversity: Cover easy/medium/hard, edge cases, and different sub-categories.
• Quality: Reference answers must be gold-standard correct. A wrong reference penalizes correct outputs.

Interpreting Results

With 50+ eval examples, a difference of ~0.3 points (on 1–5 scale) is usually meaningful.

Evaluating RFT Models

1. Evaluate with a DIFFERENT rubric than the training grader — otherwise you measure overfitting to the grader.
2. Use F1ScoreEvaluator for exact-match accuracy.
3. Use SimilarityEvaluator to catch semantically correct but differently formatted answers.
4. Compare against the base model, not just other fine-tunes.

Reference

• Azure AI Evaluation SDK docs
• Evaluation samples